What changed, and what a mark means
Anthropic’s November 3 update extends text watermarking to Claude Mythos 5, Opus 4.7, and Sonnet 4.6 on supported services. A watermark is an invisible signal that software can look for in the text. [1] [2]
Coverage on Amazon’s Bedrock cloud service is still rolling out, with completion due by November 10. Anthropic’s text detector remains a private preview for eligible organizations. [2]
For example, your own customer email could gain a mark after Claude proofreads it. A mark cannot prove who wrote the original or whether its claims are correct. A missing mark is not proof of human authorship. [2]
A developer setting can affect the signal
Anthropic says the mark adds no hidden characters or information about you or your organization. Apps can handle requests and replies as before. [3]
For developers using Sonnet 4.6, sampling settings control how a model chooses its next words. Non-default values can make the mark unreliable. If relying on it, Anthropic says to leave temperature, top_p, and top_k unset, including any values a software library adds. [3]
Our take
Keep the original draft, the edits, and your review notes together. If a customer asks about an email, that record gives useful context about your contribution. Check the numbers and commitments yourself; a marker cannot do that review for you.